A free agent skill that builds the minimum sufficient context a coding agent needs before implementation starts — and refuses to hand over a pack that's starved of the constraints that matter or drowned in ones that don't. Works with AI-DLC, Scrum, Kanban, Jira, GitHub Issues, Shape Up, or no formal methodology at all.
What this is: a context-engineering primitive for AI-assisted software development. It's packaged around Viksya's AI-DLC line — you'll see AI-DLC vocabulary (Intent, Bolt) in a few places, and it links to the AI-DLC Context Engineering Playbook for teams that use AI-DLC — but the procedure itself (a six-section pack, a starvation/drowning quality gate) doesn't require AI-DLC or any specific methodology.
Agents fail on defined work for one of two context reasons: starved (given only the intent, they invent assumptions and drift from real requirements) or drowned (given the whole repo and every document, they burn attention on irrelevant material and still miss the constraint that mattered). This skill is a repeatable procedure for hitting the middle: gather what's needed, write it as one pack, audit it in both directions, and don't proceed until it passes.
It owns one layer: what the agent is handed before execution starts. It doesn't run tasks, generate code, or manage sessions — it's built to sit alongside whatever orchestration or workflow tooling you already use (AI-DLC's aidlc-workflows, your ticketing system, or nothing formal at all), not replace it.
| File | Purpose |
|---|---|
SKILL.md |
The procedure and quality gate — load this into your agent harness. |
templates/context-pack.md |
Empty, copy-paste template with per-section guidance comments. |
examples/example-context-pack.md |
A full worked run — Intent through PASS verdict, two minutes to read. |
LICENSE |
Full license terms — free to use and share with attribution, not for resale or white-label. |
CHANGELOG.md |
What changed between versions. |
Claude Code: drop this folder into .claude/skills/context-pack-builder/ (project-level) or ~/.claude/skills/context-pack-builder/ (user-level). It's plain markdown with no dependencies — nothing to build or configure.
Other agent harnesses: the skill is portable by design — no scripts, no tool dependencies. Paste SKILL.md into your system prompt, project instructions, or equivalent "standing context" mechanism, and point the agent at templates/context-pack.md when it needs the blank template.
No harness at all: works as a manual checklist. Read SKILL.md, fill in templates/context-pack.md by hand, run the quality-gate checks yourself before handing the pack to whatever agent you're using.
This skill produces one artifact — a per-task context pack — and the procedure for building and auditing it. It does not cover:
- Standing, org-level context (architecture, conventions, domain, product, design documentation that persists across every task, not just one) — that's a deeper discipline than a single-pack assembly step, covered in the AI-DLC Context Engineering Playbook.
- Task execution, code generation, or session orchestration.
- Whether the task itself is worth doing.
If you hit a hard case this skill doesn't resolve — multi-team context, legacy systems with conflicting constraints, or you're ready to stand up the standing artifacts it references — the Playbook is the deeper instrument this skill is deliberately kept thin against.
The pack's verdict frontmatter field (pending / pass / fail) exists so something other than the chat transcript can check whether the gate passed before implementation is allowed to start. This skill doesn't ship that enforcement — it stays plain markdown with no dependencies — but here are two starting points to adapt.
In Claude Code — a PreToolUse hook. Configure a hook that blocks Edit/Write calls unless a pack for the current work shows verdict: pass, in .claude/settings.json:
{
"hooks": {
"PreToolUse": [
{
"matcher": "Edit|Write",
"hooks": [
{ "type": "command", "command": ".claude/hooks/require-context-pack-pass.sh" }
]
}
]
}
}#!/usr/bin/env bash
# .claude/hooks/require-context-pack-pass.sh
# Exit 2 blocks the tool call and shows stderr to the agent as the reason;
# exit 0 allows it. Adapt PACK_DIR and the scope match to your layout —
# this starting point checks for ANY passing pack, not the one for the
# specific task in progress.
PACK_DIR="${CONTEXT_PACK_DIR:-context-packs}"
if ! grep -rl "^verdict: pass" "$PACK_DIR"/context-pack.*.md >/dev/null 2>&1; then
echo "No context-pack.md with verdict: pass — run the context-pack-builder skill first." >&2
exit 2
fi
exit 0Claude Code's hook contract can change between versions — check the current hooks documentation if this doesn't match what you see.
In CI — a PR-level fallback. Catches anything that slipped past a harness without hook support. Adapt to your CI provider; this is GitHub-Actions-flavored:
- name: Require a passing context pack
run: |
if git diff --name-only "origin/${{ github.base_ref }}"... | grep -vE '^context-packs/|\.md$' | grep -q .; then
if ! grep -rl "^verdict: pass" context-packs/*.md >/dev/null 2>&1; then
echo "Source changes with no passing context pack. See context-pack-builder skill." >&2
exit 1
fi
fiBoth examples are deliberately minimal — matching on "a pack somewhere passes," not "the pack for this specific task passes." Tightening that (matching scope against the branch, ticket, or Bolt ID) is a few extra lines once you've picked a convention; it's left out here so the skill itself stays opinion-free about your branching and ticketing model.
Published free by Viksya (viksya.com) — decision frameworks for AI transformation and AI-DLC adoption.
AI-DLC (AI-Driven Development Life Cycle) is a methodology defined by AWS. Viksya is not affiliated with or endorsed by AWS.
Hard cases (multi-team, legacy, conflicting constraints) → Context Engineering Playbook. Starting adoption → AI-DLC Starter Kit. Not sure you're ready? → AI-DLC Adoption Readiness Assessment.
Free to use and share with this notice intact. Not for resale or white-label use.